Biotech has a distribution problem

My best friend, at 20 years old, has two genetic diseases with no cure.

It’s a big part of why I got into biotech.

The deeper I’ve gone, the more something has frustrated me. There’s an enormous amount of promising work in biotech, but it doesn’t spread.

Results get published, tools get built, and then they just… sit there.

They don’t reliably reach the people who could use them or push them further.

It’s tempting to explain this away with familiar answers: funding constraints, regulation, or the fact that biology is genuinely hard.

All of that is true.
But I keep coming back to something simpler:

Biotech has a distribution problem.

I mean distribution in two senses:

  1. The distribution of participation: the pathways for outsiders to take on meaningful work in biotech are too narrow.

  2. The distribution of work: valuable biotech research and industry progress doesn’t travel far beyond insular communities.

This matters because progress in biotech doesn’t just depend on discovery. It depends on compounding.

When work is legible and travels, other scientists can build on it. Adjacent fields can connect their insights to it. Talented people can find it. Funders can understand it.

For people like my friend, poor distribution isn’t abstract—it’s the reason promising research on his conditions might be sitting unused in a lab somewhere. It’s why the AI researcher who could accelerate that work never sees it. It’s why potential collaborators can’t find each other. Every failure of distribution is measurable in time: time before the right people connect, time before ideas compound, time before someone gets treatment.

This post is about why biotech doesn’t travel, why that matters, and how poor distribution blocks both frontier AI talent and younger contributors—and what we can do about it.

Why Biotech Doesn’t Travel

In software, when something gains traction, distribution can compound rapidly. People try it, share it, remix it, complain about it, and recruit themselves into your orbit.

Biotech is the opposite. Results get written up for insiders. Product updates are shaped by IP concerns, and compliance requirements. News comes filtered through a handful of academic press cycles, venture capitalists making posts, or niche communities.

This creates a paradox: some of the most important work on Earth is functionally invisible to the people who could accelerate it. That invisibility shapes who gets the chance to participate.

Everything that follows is a downstream consequence of this.

Biotech Needs Frontier AI Talent

Biotech oscillates between two modes: intuition-driven lab cycles with low information density, and what I call the “silicon maximalist” approach, which tries to simulate everything in silicon before we have the data, compute, or models to do it right.

The near future of biotech is neither. It’s closed-loop:

High-throughput experiments are designed to extract maximal information. Machine-learning models propose the next experiments. Labs execute them. An inference layer turns raw data into learnable signal. The results feed back into the model, the model updates, and the cycle repeats. Humans guide the system by choosing objectives, constraints, and exploration strategies.

If you’re familiar with reinforcement learning, this looks like learning through interaction with the real world, not training once on a static dataset. I think of this as natural RL.

I’ve noticed a lot of computational biology treats problems primarily as label problems. This has been useful, but it quietly assumes the labels we’re collecting are the right compression of reality, and that they stay meaningful across contexts.

In biology, this often fails. Labels are proxies. The hard part is noticing when you’re optimizing the proxy instead of what you actually care about, or when the thing that actually matters isn’t in your measurement set at all. The biggest wins will come from loops that help us discover better targets, better measurements, and better proxies, not just better labels.

This is exactly where frontier AI skill matters.

Frontier AI people are trained to design feedback loops instead of static pipelines – and to ask what signal is missing, not just how to fit what already exists.

The problem is distribution.

Most frontier AI people never see this work. And if they do, it’s often too opaque to understand how they could contribute. When some manage to enter, they encounter narrow gates and an implicit message: you matter only after years of traditional bio credentialing.

As a result, many never join biotech at all. They build trading systems, recommendation engines, LLMs, and defense tech - all places where tight feedback loops and fast iteration are rewarded.

The opportunity cost is enormous. Biotech forfeits the people best equipped to turn experimentation into a learning system rather than a sequence of isolated studies.

As a side note, I’m a massive fan of JURA bio’s approach here: manufacturing-aware design where AI & biology are deeply coupled.

Biotech Needs More Young People

Biotech has a permissioning problem.

In biotech, the default filter is credentials, not output. Some of this is completely justified: Biology is hard. Mistakes can hurt people. Regulatory environments are complex. But this framing ignores opportunity cost.

Delays hurt people too.

If it takes a decade just to be allowed into the room, the room selects for a narrow population: people who can afford long credential marathons, tolerate years of low pay, and delay risk until their lives make it harder to take any – i.e. you get a mortgage, have kids, and suddenly a moonshot feels incredibly irresponsible.

This is measurable: the median time from finishing a PhD to receiving a first major NCI research grant doubled from 6 years in 1990 to 12 years in 2016. That’s twelve years before most academic researchers gain real independence to pursue their own directions.

Historically, biotech has responded to this with: biology is too complex for newcomers.

This is partly true. But the learning curve has changed. Papers, textbooks, and whole subfields are now easier to digest with LLMs. More importantly, there are many ways to contribute that don’t require you to run clinical trials: tooling, automation, measurement, data infrastructure, literature mapping, evaluation loops, etc.

If we require a decade of credentialing before someone can matter, we select against the people most willing to try new approaches.

So What Can We Do About It?

Biotech in Plain English

Imagine sending a biotech paper to a smart friend outside the field, and they’re just as excited as you are.

Example:

Instead of:

During training, we apply a noising schedule over timesteps T that gradually corrupts antibody residue frames toward a noise distribution. We add 3D Gaussian noise to the C-alpha translations and apply Brownian motion on the space of 3D rotations for the orientations. We then train RFdiffusion to predict the denoised structure at each timestep by minimizing the mean squared error between the clean frames X0 and the network’s predicted frames. (this is a real article, here’s the link: PMC)

Perhaps something like this:

During training, we intentionally add controlled random shifts and rotations to the antibody backbone and teach the model to remove that corruption one step at a time. At generation time, this learned “reverse process” lets the model start from noise and iteratively refine it into a realistic antibody structure consistent with the constraints we provide.

Imagine if every paper, every product launch, every company update had two versions:

  • the full technical version, and
  • a plain-English version that explains what changed, why it matters, and what it enables.

When people can understand something, they can support it, join it, fund it, argue about it, and improve it.

EDIT: While writing this, I realized something like this should exist - quickly spun this up.

Build/Research in Public (with Less Jargon)

I really appreciated this tweet from Dr. Shelby on biotech’s lack of narrative. Here’s my own spin on it:

Biotech work is fragmented across silos. Researchers often don’t see what adjacent labs or subfields are doing.

We need more low-friction connection points: online spaces where people share work-in-progress in simplified language.

The simplified language part matters. Biology is huge, and the “jargon surface area” is basically infinite.

And to everyone already writing and podcasting about biotech: thank you.

Default to curiosity before critique.

In tech, when someone shows you something new, the default reaction is curiosity: “If this worked, what could it unlock?”

In biotech, the default is critique: “Here are all the reasons this won’t work.”

Critique is important, but it can’t be the only response. Curiosity creates space for orthogonal approaches, especially from outsiders who don’t share the field’s assumptions.

Prioritize more frontier AI

Data pipelines in many labs are built to analyze results after experiments run, not to systematically decide what should happen next.

Computational biology has historically been forced into this mode because determining what should happen next was largely a biology-heavy decision, dependent on intuition and slow trial-and-error in the lab.

AI changes that.

For the first time, we can formalize this question. We can model uncertainty explicitly, estimate information gain, and use models to propose experiments based on how much they improve downstream decisions… not just how well they explain existing data.

Frontier AI people are trained to think this way: as designers of systems that learn through interaction. Systems where each experiment is chosen to make the next decision better.

My goal is to bring that mindset into biotech. To help build workflows where wet labs and models are tightly coupled, uncertainty is more explicit, and experimentation itself is treated as something you can optimize.

The shift I’m aiming for is subtle but important: moving from “What does this data say?” to “What experiment would most improve our ability to act?”

The next era of biotech will come from teams that:

  • combine frontiers across domains,

  • make their work legible enough to travel,

  • and build real on-ramps for people who want to contribute now, not in ten years.

Biotech doesn’t just need better science. It needs better distribution.

I want to hear what you think and what you’re building. DM me!